Convolutional Neural Networks¶
Project: Write an Algorithm for a Dog Identification App¶
In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!
Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the Jupyter Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.
In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.
Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.
The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this Jupyter notebook.
Why We're Here¶
In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).
In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!
The Road Ahead¶
We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.
- Step 0: Import Datasets
- Step 1: Detect Humans
- Step 2: Detect Dogs
- Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
- Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)
- Step 5: Write your Algorithm
- Step 6: Test Your Algorithm
Step 0: Import Datasets¶
Make sure that you've downloaded the required human and dog datasets:
Download the dog dataset. Unzip the folder and place it in this project's home directory, at the location
/dogImages.Download the human dataset. Unzip the folder and place it in the home directory, at location
/lfw.
Note: If you are using a Windows machine, you are encouraged to use 7zip to extract the folder.
In the code cell below, we save the file paths for both the human (LFW) dataset and dog dataset in the numpy arrays human_files and dog_files.
import os
from pathlib import Path
from typing import Callable as func
import cv2
from glob import glob
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
from PIL import Image
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import torch
import torch.nn as nn
import torch.optim as optim
from torch.nn import functional as F
import torch.utils.data as data
import torchvision
from torchvision import models as models
from torchvision import transforms as T
from tqdm import tqdm
available_hardware = "cuda" if torch.cuda.is_available() else "cpu"
device = torch.device(available_hardware)
use_cuda = torch.cuda.is_available()
# load filenames for human and dog images
human_files = np.array(glob("lfw/*/*"))
dog_files = np.array(glob("dogImages/*/*/*"))
# print number of images in each dataset
print('There are %d total human images.' % len(human_files))
print('There are %d total dog images.' % len(dog_files))
Step 1: Detect Humans¶
In this section, we use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images.
OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory. In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.
# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')
# load color (BGR) image
img = cv2.imread(human_files[1800])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# find faces in image
faces = face_cascade.detectMultiScale(gray)
# print number of faces detected in the image
print('Number of faces detected:', len(faces))
# get bounding box for each detected face
for (x,y,w,h) in faces:
# add bounding box to color image
cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()
Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.
In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.
Write a Human Face Detector¶
We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
img = cv2.imread(img_path)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
faces = face_cascade.detectMultiScale(gray)
return len(faces) > 0
(IMPLEMENTATION) Assess the Human Face Detector¶
Question 1: Use the code cell below to test the performance of the face_detector function.
- What percentage of the first 100 images in
human_fileshave a detected human face? - What percentage of the first 100 images in
dog_fileshave a detected human face?
Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.
Answer: (You can print out your results and/or write your percentages in this cell)
human_files_short = human_files[:100]
dog_files_short = dog_files[:100]
#-#-# Do NOT modify the code above this line. #-#-#
## TODO: Test the performance of the face_detector algorithm
## on the images in human_files_short and dog_files_short.
def calculate_human_detection_percentage(images: list, images_name: str) -> str:
detection_results = list(map(face_detector, images))
success_count = sum(map(int, detection_results))
success_percentage = success_count / len(detection_results) * 100
result_message = f"Human face was detected in {success_percentage}% of {images_name}."
return result_message
print(calculate_human_detection_percentage(human_files_short, "human pictures"))
print(calculate_human_detection_percentage(dog_files_short, "dog pictures"))
We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.
### (Optional)
### TODO: Test performance of another face detection algorithm.
### Feel free to use as many code cells as needed.
class ImageDataSet(data.Dataset):
def __init__(self, human_paths: list, dog_paths: list) -> None:
self.paths = self.__make_path_list(human_paths, dog_paths)
self.transform = self.__make_transform()
@staticmethod
def __make_label(human: bool) -> torch.Tensor:
return torch.Tensor((int(human), int(not human)))
@staticmethod
def __make_path_list(human_paths: list, dog_paths: list) -> list:
paths = []
for human_path in human_paths:
paths.append((human_path, ImageDataSet.__make_label(human=True)))
for dog_path in dog_paths:
paths.append((dog_path, ImageDataSet.__make_label(human=False)))
return paths
@staticmethod
def __make_transform():
channel_means = [0.485, 0.456, 0.406]
channel_stds = [0.229, 0.224, 0.225]
return T.Compose([T.Resize((224,224)),
T.ToTensor(),
T.Normalize(mean=channel_means, std=channel_stds)])
def __len__(self) -> int:
return len(self.paths)
def __getitem__(self, index) -> torch.Tensor:
image_path, label = self.paths[index]
pil_image = Image.open(image_path)
torch_image = self.transform(pil_image)
return torch_image, label
human_validation_index = int(len(human_files[100:]) * 0.8)
dog_validation_index = int(len(dog_files[100:]) * 0.8)
human_train_paths = human_files[100:human_validation_index]
human_validation_paths = human_files[human_validation_index:]
dog_train_paths = dog_files[100:dog_validation_index]
dog_validation_paths = dog_files[dog_validation_index:]
batch_size = 64
train_data = ImageDataSet(human_train_paths, dog_train_paths)
train_loader = data.DataLoader(train_data, batch_size=batch_size, shuffle=True, pin_memory=True)
validation_data = ImageDataSet(human_validation_paths, dog_validation_paths)
validation_loader = data.DataLoader(validation_data, batch_size=batch_size, shuffle=True, pin_memory=True)
class FaceDetectorMlp(nn.Module):
def __init__(self):
super().__init__()
self.hidden = nn.Linear(2048, 1024)
self.output = nn.Linear(1024, 2)
self.dropout = nn.Dropout(p=0.2)
def forward(self, input):
hidden_out = self.dropout(F.relu(self.hidden(input)))
output_out = torch.sigmoid(self.output(hidden_out))
return output_out
def load_resnet() -> nn.Module:
resnet = torchvision.models.resnet50(pretrained=True)
resnet.to(device)
for module in resnet.modules():
for param in module.parameters():
param.requires_grad = False
return resnet
def build_face_detector() -> nn.Module:
frozen_resnet = load_resnet()
face_detector = FaceDetectorMlp().to(device)
frozen_resnet.fc = face_detector
return frozen_resnet
def test_network_architecture(network: nn.Module) -> None:
modules = dict(network.named_children())
for name in modules:
if name != "fc":
module = modules[name]
for param in module.parameters():
assert not param.requires_grad
for param in modules["fc"].parameters():
assert param.requires_grad
network = build_face_detector()
test_network_architecture(network)
optimizer = torch.optim.Adam(network.fc.parameters(), lr=0.001)
criterion = torch.nn.BCELoss()
epoch_count = 1
def training_step(images: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
optimizer.zero_grad()
output = network.forward(images.to(device))
loss = criterion(output, labels.to(device))
loss.backward()
optimizer.step()
return loss
def train() -> None:
network.train()
train_loss = 0
for images, labels in train_loader:
train_loss += training_step(images, labels)
train_loss /= len(train_loader)
print(f"Averge training loss: {train_loss}")
def validation_step(images: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
output = network.forward(images.to(device))
loss = criterion(output, labels.to(device))
return loss
def validate() -> None:
network.eval()
with torch.no_grad():
validation_loss = 0
for images, labels in validation_loader:
validation_loss += validation_step(images, labels)
validation_loss /= len(validation_loader)
print(f"Average validation loss: {validation_loss}")
for epoch in range(epoch_count):
print(f"Epoch {epoch}")
train()
validate()
def make_imagenet_transform():
expected_image_size = (224, 224)
channel_means = [0.485, 0.456, 0.406]
channel_stds = [0.229, 0.224, 0.225]
transform = T.Compose([T.Resize(expected_image_size),
T.ToTensor(),
T.Normalize(mean=channel_means, std=channel_stds)])
return transform
def load_image_as_minibatch(path: str) -> torch.Tensor:
pil_image = Image.open(path)
transform = make_imagenet_transform()
torch_image = transform(pil_image)
image_batch = torch_image.unsqueeze(0)
return image_batch
def is_face_detected(image_path: str) -> bool:
torch_image = load_image_as_minibatch(image_path)
output = network(torch_image.to(device))
flat_output = output.view(-1)
face_probability = flat_output[0].item()
dog_probability = flat_output[1].item()
return face_probability > dog_probability
def print_detection_percentage(images: list, detector: func, object_name: str, images_name: str) -> None:
detection_results = list(map(detector, images))
success_count = sum(map(int, detection_results))
success_percentage = success_count / len(detection_results) * 100
result_message = f"{object_name} was detected in {success_percentage}% of {images_name}."
print(result_message)
network.eval()
print_detection_percentage(human_files_short, is_face_detected, "Human face", "human pictures")
print_detection_percentage(dog_files_short, is_face_detected, "Human face", "dog pictures")
torch.save(network.state_dict(), "resnet_face_detector.pt")
Step 2: Detect Dogs¶
In this section, we use a pre-trained model to detect dogs in images.
Obtain Pre-trained VGG-16 Model¶
The code cell below downloads the VGG-16 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories.
# define VGG16 model
VGG16 = models.vgg16(pretrained=True)
# check if CUDA is available
use_cuda = torch.cuda.is_available()
# move model to GPU if CUDA is available
if use_cuda:
VGG16 = VGG16.cuda()
Given an image, this pre-trained VGG-16 model returns a prediction (derived from the 1000 possible categories in ImageNet) for the object that is contained in the image.
(IMPLEMENTATION) Making Predictions with a Pre-trained Model¶
In the next code cell, you will write a function that accepts a path to an image (such as 'dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg') as input and returns the index corresponding to the ImageNet class that is predicted by the pre-trained VGG-16 model. The output should always be an integer between 0 and 999, inclusive.
Before writing the function, make sure that you take the time to learn how to appropriately pre-process tensors for pre-trained models in the PyTorch documentation.
def get_prediction(image_path: str, network: nn.Module) -> int:
image_batch = load_image_as_minibatch(image_path)
network.eval()
output = network.forward(image_batch.to(device))
probabilities = F.softmax(output, dim=1)
top_prob, top_index = torch.topk(probabilities, k=1)
return top_index.item()
def VGG16_predict(img_path: str) -> int:
'''
Use pre-trained VGG-16 model to obtain index corresponding to
predicted ImageNet class for image at specified path
Args:
img_path: path to an image
Returns:
Index corresponding to VGG-16 model's prediction
'''
## TODO: Complete the function.
## Load and pre-process an image from the given img_path
## Return the *index* of the predicted class for that image
return get_prediction(img_path, VGG16)
VGG16_predict(dog_files_short[50])
(IMPLEMENTATION) Write a Dog Detector¶
While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained VGG-16 model, we need only check if the pre-trained model predicts an index between 151 and 268 (inclusive).
Use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
## TODO: Complete the function.
prediction_index = VGG16_predict(img_path)
first_dog_index = 151
last_dog_index = 268
is_index_dog = prediction_index in range(first_dog_index, last_dog_index+1)
return is_index_dog
dog_detector(dog_files_short[16])
(IMPLEMENTATION) Assess the Dog Detector¶
Question 2: Use the code cell below to test the performance of your dog_detector function.
- What percentage of the images in
human_files_shorthave a detected dog? - What percentage of the images in
dog_files_shorthave a detected dog?
Answer:
### TODO: Test the performance of the dog_detector function
### on the images in human_files_short and dog_files_short.
print("Using VGG-16")
print_detection_percentage(human_files_short, dog_detector, "A dog", "human_files_short")
print_detection_percentage(dog_files_short, dog_detector, "A dog", "dog_files_short")
We suggest VGG-16 as a potential network to detect dog images in your algorithm, but you are free to explore other pre-trained networks (such as Inception-v3, ResNet-50, etc). Please use the code cell below to test other pre-trained PyTorch models. If you decide to pursue this optional task, report performance on human_files_short and dog_files_short.
### (Optional)
### TODO: Report the performance of another pre-trained network.
### Feel free to use as many code cells as needed.
resnet50 = models.resnet50(pretrained=True)
resnet50.to(device)
def dog_detector_resnet(img_path):
prediction_index = get_prediction(img_path, resnet50)
first_dog_index = 151
last_dog_index = 268
is_index_dog = prediction_index in range(first_dog_index, last_dog_index+1)
return is_index_dog
print("Using ResNet50")
print_detection_percentage(human_files_short, dog_detector_resnet, "A dog", "human_files_short")
print_detection_percentage(dog_files_short, dog_detector_resnet, "A dog", "dog_files_short")
Step 3: Create a CNN to Classify Dog Breeds (from Scratch)¶
Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 10%. In Step 4 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.
We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have trouble distinguishing between a Brittany and a Welsh Springer Spaniel.
| Brittany | Welsh Springer Spaniel |
|---|---|
 |
 |
It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).
| Curly-Coated Retriever | American Water Spaniel |
|---|---|
 |
 |
Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.
| Yellow Labrador | Chocolate Labrador | Black Labrador |
|---|---|---|
 |
 |
 |
We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.
Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!
(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset¶
Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dogImages/train, dogImages/valid, and dogImages/test, respectively). You may find this documentation on custom datasets to be a useful resource. If you are interested in augmenting your training and/or validation data, check out the wide variety of transforms!
### TODO: Write data loaders for training, validation, and test sets
## Specify appropriate transforms, and batch_sizes
def make_loader(image_dir: str, transform) -> data.DataLoader:
batch_size = 24
data_set = torchvision.datasets.ImageFolder(image_dir, transform=transform)
loader = data.DataLoader(data_set, batch_size, shuffle=True, pin_memory=True)
return loader
def find_channel_means_and_stds(image_dir: str) -> torch.Tensor:
resize_transform = T.Compose([T.Resize((224, 224)), T.ToTensor()])
loader = make_loader(image_dir, resize_transform)
means = torch.zeros(3).to(device)
stds = torch.zeros(3).to(device)
for images, _labels in loader:
flat_images = images.to(device).flatten(2)
means += torch.mean(flat_images, dim=[0,2])
stds += torch.std(flat_images, dim=[0,2])
means /= len(loader)
stds /= len(loader)
return torch.stack((means, stds))
def find_channel_mean_and_std_for_all_images() -> (torch.Tensor, torch.Tensor):
train_stats = find_channel_means_and_stds("dogImages/train")
valid_stats = find_channel_means_and_stds("dogImages/valid")
test_stats = find_channel_means_and_stds("dogImages/test")
total_stats = torch.mean(torch.stack((train_stats, valid_stats, test_stats)), dim=0)
return total_stats[0], total_stats[1]
# dog_means, dog_stds = find_channel_mean_and_std_for_all_images()
# Hardcoding the values below to avoid repeating the calculations
dog_means = [0.4887, 0.4678, 0.3988]
dog_stds = [0.2586, 0.2530, 0.2584]
print(dog_means)
print(dog_stds)
image_size = 224
def make_train_transform():
return T.Compose([T.RandomRotation(15),
T.RandomHorizontalFlip(),
T.Resize(232),
T.RandomCrop(image_size),
T.ToTensor(),
T.Normalize(mean=dog_means, std=dog_stds)])
def make_eval_transform():
return T.Compose([T.Resize(232),
T.CenterCrop(image_size),
T.ToTensor(),
T.Normalize(mean=dog_means, std=dog_stds)])
def make_loader_dict() -> dict:
loader_dict = {}
loader_dict["train"] = make_loader("dogImages/train", make_train_transform())
loader_dict["valid"] = make_loader("dogImages/valid", make_eval_transform())
loader_dict["test"] = make_loader("dogImages/test", make_eval_transform())
return loader_dict
loader_dict = make_loader_dict()
images, labels = next(iter(loader_dict["train"]))
images.shape
Question 3: Describe your chosen procedure for preprocessing the data.
- How does your code resize the images (by cropping, stretching, etc)? What size did you pick for the input tensor, and why?
- Did you decide to augment the dataset? If so, how (through translations, flips, rotations, etc)? If not, why not?
Answer:
- For the training set, images are resized so that the smaller edge is 232 px and then randomly cropped to a 224x224 px square. This should eliminate some of the background pixels near the edges, increasing density of relevant features in the image. The randomized crop position should help develop better translation invariance. It might also improve detection of partially obscured objects (dogs in this case), due to some parts of the object being cropped-out sometimes. Validation and test images are resized to 232 px and center-cropped to 224x224 px without randomization. The 224 px size was chosen based on fact it was also used for the pre-trained PyTorch models.
- Images from all 3 sets are normalized using average color channel means and standard deviations, calculated accross all available dog images. Additionally, training images have random rotation, horizontal flip and random crop applied, to help the model develop better invariances.
(IMPLEMENTATION) Model Architecture¶
Create a CNN to classify dog breed. Use the template in the code cell below.
class_dirs = [dir for dir in os.scandir("dogImages/test") if dir.is_dir()]
class_count = len(class_dirs)
class_count
# define the CNN architecture
class ConvModule(nn.Module):
def __init__(self, input_depth: int, conv_depths: list, pool_kernel: int, pool_stride: int) -> None:
super().__init__()
assert len(conv_depths) > 0, "At least 1 conv_depth must be provided."
self.conv_layers = nn.ModuleList()
depths = [input_depth]
depths.extend(conv_depths)
for i in range(1,len(depths)):
conv = nn.Conv2d(depths[i-1], depths[i], kernel_size=3, padding=1)
self.conv_layers.append(conv)
self.pool = nn.MaxPool2d(kernel_size=pool_kernel, stride=pool_stride)
def forward(self, input):
conv_out = input
for conv in self.conv_layers:
conv_out = F.relu(conv(conv_out))
output = self.pool(conv_out)
return output
class Net(nn.Module):
### TODO: choose an architecture, and complete the class
def __init__(self):
super().__init__()
## Define layers of a CNN
self.conv1 = ConvModule(input_depth=3, conv_depths=[64,64,64], pool_kernel=2, pool_stride=2)
self.conv2 = ConvModule(input_depth=64, conv_depths=[128,128,128], pool_kernel=2, pool_stride=2)
self.conv3 = ConvModule(input_depth=128, conv_depths=[256,256,256], pool_kernel=2, pool_stride=2)
self.conv4 = ConvModule(input_depth=256, conv_depths=[512,512,512], pool_kernel=2, pool_stride=2)
self.fc = nn.Linear(14*14*512+8*8*128, class_count)
self.sampler = nn.AvgPool2d(kernel_size=7, stride=7)
self.dropout = nn.Dropout2d(p=0.5)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
sample = self.dropout(self.sampler(x))
x = self.conv3(x)
x = self.conv4(x)
sample = sample.view(sample.shape[0], -1)
x = x.view(x.shape[0], -1)
x = torch.cat([x, sample], dim=1)
x = F.log_softmax(self.fc(x), dim=1)
return x
#-#-# You do NOT have to modify the code below this line. #-#-#
# instantiate the CNN
model_scratch = Net()
# move tensors to GPU if CUDA is available
if use_cuda:
model_scratch.cuda()
Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step.
Answer:
- First idea was to use one of the most successful architectures, such VGG-16 or ResNet50. However, training such deep networks would require too much computational resources to be feasible for this project.
- A simple CNN architecture, inspired by the previous Udacity projects was designed. It consisted of a stack of Conv2d and MaxPool2d layers, ending with 2 fully-connected layers. It was small enough to be trained on local hardware. The layers were grouped into modules implemented by the ConvModule class. Each of these modules consisted of 1 or more conv2d layers, which detect image features of increasing complexity and one final max pooling layer, which reduces the x-y dimensions of feature maps. This reduction is important for decreasing memory requirements and for eliminating redundant information from the feature maps.
- Several short training runs were performed to experiment with the architecture. Adjustments to conv layer depth and count were made, eventually reaching ~8% accuracy after 10 epochs of training.
- A new idea was implemented: taking the output of one of the middle conv layers, pooling it, and plugging it directly into the final FC layer, after being combined with output of the final conv layer. This modification led to a marked improvement, enabling the network to reach ~15% accuracy.
- To reduce the observed divergence of training and validation loss, dropout was added for the middle conv output. After some tuning, this eventually further improved the accuracy to around 17% (after 10 epochs).
- Since the process mentioned above took several days, it was decided at that point to stop further experimentation and to settle with the current architecture. For the final training session, the model was extended with additional layers and then trained for 100 epochs overnight in the Udacity GPU environment. It managed to reach 29% test accuracy.
(IMPLEMENTATION) Specify Loss Function and Optimizer¶
Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_scratch, and the optimizer as optimizer_scratch below.
### TODO: select loss function
criterion_scratch = nn.NLLLoss()
### TODO: select optimizer
optimizer_scratch = optim.Adam(model_scratch.parameters(), lr=0.001)
(IMPLEMENTATION) Train and Validate the Model¶
Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_scratch.pt'.
torch.cuda.empty_cache()
save_file = 'model_scratch.pt'
def save_model(model: nn.Module, file_path: str) -> None:
model_state = model.state_dict()
torch.save(model_state, file_path)
print(f"Model state saved to {file_path}")
def train(n_epochs, loaders, model, optimizer, criterion, use_cuda, save_path):
"""returns trained model"""
# initialize tracker for minimum validation loss
valid_loss_min = np.Inf
for epoch in range(1, n_epochs+1):
print(f"Epoch {epoch}")
# initialize variables to monitor training and validation loss
train_loss = 0.0
valid_loss = 0.0
###################
# train the model #
###################
model.train()
for data, target in loaders['train']:
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda()
optimizer.zero_grad()
output = model.forward(data)
batch_loss = criterion(output, target)
batch_loss.backward()
optimizer.step()
train_loss += batch_loss
train_loss /= len(loaders["train"])
######################
# validate the model #
######################
model.eval()
for data, target in loaders['valid']:
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda()
with torch.no_grad():
output = model.forward(data)
batch_loss = criterion(output, target)
valid_loss += batch_loss
valid_loss /= len(loaders["valid"])
# print training/validation statistics
print('Training Loss: {:.6f} \tValidation Loss: {:.6f}'.format(train_loss, valid_loss))
## TODO: save the model if validation loss has decreased
if valid_loss < valid_loss_min:
save_model(model, save_path)
valid_loss_min = valid_loss
# return trained model
return model
if Path(save_file).exists():
model_scratch.load_state_dict(torch.load(save_file))
print(f"Existing model loaded for training from {save_file}.")
# train the model
model_scratch = train(3, loader_dict, model_scratch, optimizer_scratch,
criterion_scratch, use_cuda, save_file)
# load the model that got the best validation accuracy
model_scratch.load_state_dict(torch.load(save_file))
(IMPLEMENTATION) Test the Model¶
Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 10%.
def test(loaders, model, criterion, use_cuda):
# monitor test loss and accuracy
test_loss = 0.
correct = 0.
total = 0.
model.eval()
for batch_idx, (data, target) in enumerate(loaders['test']):
# move to GPU
if use_cuda:
data, target = data.cuda(), target.cuda()
# forward pass: compute predicted outputs by passing inputs to the model
output = model(data)
# calculate the loss
loss = criterion(output, target)
# update average test loss
test_loss = test_loss + ((1 / (batch_idx + 1)) * (loss.data - test_loss))
# convert output probabilities to predicted class
pred = output.data.max(1, keepdim=True)[1]
# compare predictions to true label
correct += np.sum(np.squeeze(pred.eq(target.data.view_as(pred))).cpu().numpy())
total += data.size(0)
print('Test Loss: {:.6f}\n'.format(test_loss))
print('\nTest Accuracy: %2d%% (%2d/%2d)' % (
100. * correct / total, correct, total))
torch.cuda.empty_cache()
model_scratch.load_state_dict(torch.load(save_file))
# call test function
test(loader_dict, model_scratch, criterion_scratch, use_cuda)
Step 4: Create a CNN to Classify Dog Breeds (using Transfer Learning)¶
You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.
(IMPLEMENTATION) Specify Data Loaders for the Dog Dataset¶
Use the code cell below to write three separate data loaders for the training, validation, and test datasets of dog images (located at dogImages/train, dogImages/valid, and dogImages/test, respectively).
If you like, you are welcome to use the same data loaders from the previous step, when you created a CNN from scratch.
## TODO: Specify data loaders
loader_dict
(IMPLEMENTATION) Model Architecture¶
Use transfer learning to create a CNN to classify dog breed. Use the code cell below, and save your initialized model as the variable model_transfer.
## TODO: Specify model architecture
class BreedClassifierMlp(nn.Module):
def __init__(self):
super().__init__()
self.hidden = nn.Linear(2048, 1024)
self.output = nn.Linear(1024, class_count)
self.dropout = nn.Dropout(p=0.2)
def forward(self, input):
hidden_out = self.dropout(F.relu(self.hidden(input)))
output_out = F.log_softmax(self.output(hidden_out), dim=1)
return output_out
def load_resnet() -> nn.Module:
resnet = models.resnet50(pretrained=True)
resnet.to(device)
for module in resnet.modules():
for param in module.parameters():
param.requires_grad = False
return resnet
def build_breed_network() -> nn.Module:
frozen_resnet = load_resnet()
breed_classifier = BreedClassifierMlp().to(device)
frozen_resnet.fc = breed_classifier
return frozen_resnet
def test_network_architecture(network: nn.Module) -> None:
modules = dict(network.named_children())
for name in modules:
if name != "fc":
module = modules[name]
for param in module.parameters():
assert not param.requires_grad
for param in modules["fc"].parameters():
assert param.requires_grad
model_transfer = build_breed_network()
test_network_architecture(model_transfer)
Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.
Answer: I used similar architecutre as the one implemented for the optional human face detector earlier in the project. ResNet50 was chosen as the base CNN due to its very good performance in the ImageNet challenge.
After loading, the pre-trained model weights are frozen and then the FC layer is replaced with a custom fully-connected classifier. Finally, a test function is called to check if the feature detector's layers are still frozen, while the the classifier's are not.
The ResNet feature detection layers have been pre-trained on the ImageNet data set, which contains dog breeds among its classes. And since ResNet has reached a very high classification accuracy for that set, it can be assumed that it has learned to detect dog breed features well. To accomplish the current task, these features must now be used to classify a new set of dog breeds, which will be done by the attached custom MLP. This approach shouldn't require much training, since only the two-layer fully-connected classifier needs to be trained.
(IMPLEMENTATION) Specify Loss Function and Optimizer¶
Use the next code cell to specify a loss function and optimizer. Save the chosen loss function as criterion_transfer, and the optimizer as optimizer_transfer below.
criterion_transfer = nn.NLLLoss()
optimizer_transfer = optim.Adam(params=model_transfer.fc.parameters(), lr=0.001)
epoch_count = 10
(IMPLEMENTATION) Train and Validate the Model¶
Train and validate your model in the code cell below. Save the final model parameters at filepath 'model_transfer.pt'.
# train the model
model_transfer = train(epoch_count, loader_dict, model_transfer, optimizer_transfer, criterion_transfer, use_cuda, 'model_transfer.pt')
# load the model that got the best validation accuracy (uncomment the line below)
model_transfer.load_state_dict(torch.load('model_transfer.pt'))
(IMPLEMENTATION) Test the Model¶
Try out your model on the test dataset of dog images. Use the code cell below to calculate and print the test loss and accuracy. Ensure that your test accuracy is greater than 60%.
test(loader_dict, model_transfer, criterion_transfer, use_cuda)
(IMPLEMENTATION) Predict Dog Breed with the Model¶
Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan hound, etc) that is predicted by your model.
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.
# list of class names by index, i.e. a name can be accessed like class_names[0]
class_list = loader_dict["test"].dataset.classes
class_names = [item[4:].replace("_", " ").title() for item in class_list]
def load_dog_image(path: str) -> torch.Tensor:
pil_image = Image.open(path)
transform = loader_dict["valid"].dataset.transform
torch_image = transform(pil_image)
image_as_batch = torch_image.unsqueeze(0)
return image_as_batch
def predict_breed_transfer(img_path: str, model: nn.Module, model_state_path: str) -> str:
image = load_dog_image(img_path)
model.load_state_dict(torch.load(model_state_path))
model.to(device)
output = model.forward(image.to(device))
_top_prob, top_index = torch.topk(output, k=1)
top_dog = class_names[top_index]
return top_dog
predict_breed_transfer("dogImages/train/029.Border_collie/Border_collie_02009.jpg", model_transfer, "model_transfer.pt")
Step 5: Write your Algorithm¶
Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,
- if a dog is detected in the image, return the predicted breed.
- if a human is detected in the image, return the resembling dog breed.
- if neither is detected in the image, provide output that indicates an error.
You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 4 to predict dog breed.
Some sample output for our algorithm is provided below, but feel free to design your own user experience!
(IMPLEMENTATION) Write your Algorithm¶
### TODO: Write your algorithm.
### Feel free to use as many code cells as needed.
class DogVsMan:
def __init__(self, dog_classifier_path: str, face_detector_path: str) -> None:
self.dog_detector = self.__prepare_dog_detector()
self.dog_classifier = self.__prepare_dog_classifier(dog_classifier_path)
self.face_detector_cnn = self.__prepare_face_detector(face_detector_path)
@staticmethod
def __prepare_dog_detector() -> nn.Module:
resnet50 = models.resnet50(pretrained=True)
resnet50.to(device)
return resnet50
def __is_image_dog(self, img_path: str) -> bool:
prediction_index = get_prediction(img_path, self.dog_detector)
first_dog_index = 151
last_dog_index = 268
is_index_dog = prediction_index in range(first_dog_index, last_dog_index+1)
return is_index_dog
@staticmethod
def __prepare_face_detector(save_path: str) -> nn.Module:
face_detector = build_face_detector()
save_file = torch.load(save_path)
face_detector.load_state_dict(save_file)
face_detector.to(device)
return face_detector
@staticmethod
def __detect_face_using_cascade(image_path: str) -> bool:
image = cv2.imread(image_path)
gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
detected_faces = face_cascade.detectMultiScale(gray_image)
return len(detected_faces) > 0
def __detect_face_using_cnn(self, image_path: str) -> bool:
torch_image = load_image_as_minibatch(image_path)
output = self.face_detector_cnn(torch_image.to(device))
flat_output = output.view(-1)
face_probability = flat_output[0].item()
dog_probability = flat_output[1].item()
return face_probability > dog_probability
@staticmethod
def __prepare_dog_classifier(save_path: str) -> nn.Module:
classifier = build_breed_network()
classifier.to(device)
save_file = torch.load(save_path)
classifier.load_state_dict(save_file)
return classifier
def __classify_breed(self, image_path: str) -> (str, int):
image = load_dog_image(image_path)
output = self.dog_classifier.forward(image.to(device))
_top_prob, top_index = torch.topk(output, k=1)
breed_name = class_names[top_index]
return breed_name, top_index.item()
@staticmethod
def __show_image(path: str) -> None:
plt.axis("off")
image = plt.imread(path)
plt.imshow(image)
plt.show()
@staticmethod
def __select_dog_image_of_class(class_index: int) -> str:
dir_list = list(os.scandir("dogImages/train"))
for dir in dir_list:
if str(class_index+1) in dir.path:
class_dir = dir
break
image_path_list = list(os.scandir(class_dir))
image_count = len(image_path_list)
selected_image_index = np.random.randint(image_count)
selected_image_path = image_path_list[selected_image_index].path
return selected_image_path
def __detect_dogs_and_faces(self, image_path: str, use_cnn: bool) -> (bool, bool):
is_dog = self.__is_image_dog(image_path)
is_man = None
if use_cnn:
is_man = self.__detect_face_using_cnn(image_path)
else:
is_man = self.__detect_face_using_cascade(image_path)
return is_dog, is_man
@staticmethod
def __print_summary(is_dog: bool, is_man: bool, breed="N\A") -> None:
print("\nSUMMARY")
print("-------")
print(f"Dog detected: {is_dog} | Face detected: {is_man} | Breed classification: {breed}\n\n")
def interpret_image(self, image_path: str, use_cnn_face_detector: bool) -> None:
print(f"I'm looking at {image_path}:")
self.__show_image(image_path)
is_dog, is_man = self.__detect_dogs_and_faces(image_path, use_cnn_face_detector)
if not is_dog and not is_man:
print("I have no idea what that is!")
self.__print_summary(is_dog, is_man)
return
if is_dog and is_man:
print("This could be both a dog or a person, but there's a striking resemblance to a...")
else:
verdict = "dog" if is_dog else "person"
print(f"I think this is a {verdict}, that looks like a...")
breed, breed_index = self.__classify_breed(image_path)
breed_example = self.__select_dog_image_of_class(breed_index)
self.__show_image(breed_example)
print(f"{breed}!")
self.__print_summary(is_dog, is_man, breed)
app = DogVsMan("model_transfer.pt", "resnet_face_detector.pt")
app.interpret_image("other1.jpg", use_cnn_face_detector=False)
app.interpret_image("dog1.jpg", use_cnn_face_detector=False)
app.interpret_image("man1.jpg", use_cnn_face_detector=False)
Step 6: Test Your Algorithm¶
In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?
(IMPLEMENTATION) Test Your Algorithm on Sample Images!¶
Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.
Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.
Answer: (Three possible points for improvement) Dog breed classification works poorly. Even though the model reached around 80% accuracy for the test set, the results provided in the app are not satisfactory. Even for dog images it fails to find the correct breed. A better model should be used. Possible improvements:
- Better training data augmentation.
- Different architecture used as a base for transfer learning, such as GoogLeNet.
- Longer training with different learning rates.
- Using the cascade face detector greatly decreases the amount of false positives for face detection. Notice that in the second set of exaples below, where use_cnn_face_detector=True is used, the app detects a human face in each image. Even the one showing some eggplants ;)
## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.
images = ["dog1.jpg", "dog2.jpg", "man1.jpg", "man2.jpg", "other1.jpg", "other2.jpg"]
for image in images:
app.interpret_image(image, use_cnn_face_detector=False)
for image in images:
app.interpret_image(image, use_cnn_face_detector=True)
